Epilepsy is a chronic neurological disorder marked by abnormal neuronal activity and distinct seizure stages: interictal, preictal, ictal, and postictal. Electroencephalography (EEG) plays a key role in epilepsy research but is limited by data scarcity, inconsistent quality, and high inter-subject variability, which hinder model generalization. Existing data augmentation methods often focus on specific stages, overlooking variations across stages and subjects. To address this, we propose a hierarchical temporal generative adversarial network (HT-GAN) for cross-subject and cross-stage EEG data augmentation. HT-GAN integrates a self-attention mechanism to ensure temporal consistency and preserve global features, while adversarial training enhances data diversity and realism. Experimental results show that HT-GAN generates high-quality EEG data across all epileptic stages, effectively improving data availability and supporting robust deep learning applications in epilepsy research.